Concept Erasure on Broader Unsafe Categories
14.29Hate Residual PercentageScaleErasure
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| ScaleErasure2026.06 | 14.29 | 9.71 | 11.9 | 18.85 | 22.66 | 15.96 | 15.56 | |
| SLDthreshold=strong2026.06 | 17.32 | 14.56 | 21.96 | 20.22 | 26.87 | 14.03 | 19.16 | |
| UCE2026.06 | 18.18 | 12.38 | 23.02 | 20.97 | 25.82 | 16.64 | 19.5 | |
| SLDthreshold=medium2026.06 | 20.78 | 15.53 | 22.09 | 23.85 | 29.09 | 18.98 | 21.72 | |
| RECE2026.06 | 21.65 | 16.75 | 24.87 | 25.09 | 33.29 | 22.01 | 23.94 | |
| ESDvariant=u2026.06 | 23.38 | 21.48 | 32.14 | 28.34 | 38.79 | 26 | 28.35 | |
| ESDvariant=x2026.06 | 26.41 | 25.24 | 33.99 | 28.71 | 39.02 | 23.93 | 29.55 |